Generators: Avoid calling into runtime if operand stack is empty

Generators: Avoid calling into runtime if operand stack is empty

This patch makes yield sites save the resume continuation and context
inline.  If the operand stack is empty, we can avoid a call into the
runtime.  This also makes the SuspendJSGeneratorObject runtime function
less magical: it just has to save the operand stack and stack handlers.

This speeds up the following case by a factor of 3 or so:

  function* until(n) {
    for (var i = 0; i < n; i++)
      yield i;
  }

  function sum(iter) {
    var sum = 0;
    for (var x of iter) sum += x;
    return sum;
  }

  for (var i = 0; i < 10000; i++) sum(until(1000))

Also, there is no more sentinel value as the generators will resume in
the right place already, allowing me to remove the hack added to the
--debug-code check in r14437.

R=mstarzinger@chromium.org
BUG=

Committed: http://code.google.com/p/v8/source/detail?r=15240

[wingo, 2013-06-20 07:16:02]

Whenever the tree unfreezes, I'd like to get this one in -- it speeds up
generators considerably in the common case of an empty operand stack.

[Michael Starzinger, 2013-06-20 12:08:31]

https://codereview.chromium.org/15990004/diff/3001/src/ia32/full-codegen-ia32.cc
File src/ia32/full-codegen-ia32.cc (right):

https://codereview.chromium.org/15990004/diff/3001/src/ia32/full-codegen-ia32...
src/ia32/full-codegen-ia32.cc:1967:
Immediate(Smi::FromInt(continuation.pos())));
It's too bad that we cannot reliably predict the binding offset of "resume"
here, otherwise the two jumps at the top would be obsolete. Damn! <sarcasm>Maybe
leave a comment in that regard to eternalize our disappointment. :)</sarcasm>

https://codereview.chromium.org/15990004/diff/3001/src/ia32/full-codegen-ia32...
src/ia32/full-codegen-ia32.cc:2034: __ mov(eax, Operand(esp, (0 + 1) *
kPointerSize + handler_size));  // g
nit: The trailing '// g' comment is somewhat duplicated, let's drop it for the
'__mov' instruction.

https://codereview.chromium.org/15990004/diff/3001/src/runtime.cc
File src/runtime.cc (right):

https://codereview.chromium.org/15990004/diff/3001/src/runtime.cc#newcode2614
src/runtime.cc:2614: 
Can we leave a short comment here saying that JSGeneratorObject::context and
JSGeneratorObject::continuation are set by the caller in generated code? Maybe
even assert that both have been set correctly by the caller?

[wingo, 2013-06-20 12:45:45]

Addressed comments; PTAL.

[Michael Starzinger, 2013-06-20 12:54:05]

LGTM.

[wingo, 2013-06-20 12:59:55]

Committed patchset #2 manually as r15240 (presubmit successful).